From e4682d9cf00346e0cf8cdd5fd9f6a49c40ae13c1 Mon Sep 17 00:00:00 2001 From: Yorke Rhodes III Date: Sun, 11 Oct 2026 14:15:21 -0400 Subject: [PATCH] docs: add bounded decision-model recommendation Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --- DECISION-MODEL-RECO.md | 43 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 43 insertions(+) create mode 100644 DECISION-MODEL-RECO.md diff --git a/DECISION-MODEL-RECO.md b/DECISION-MODEL-RECO.md new file mode 100644 index 0000000..0c2264a --- /dev/null +++ b/DECISION-MODEL-RECO.md @@ -0,0 +1,43 @@ +# Decision-model recommendation: curator metadata triage only + +Date: 2026-10-11. Public-source proposal; no measured integration. +Reviewed [README](https://github.com/Ethical-Tech-CoLab/turnstone/blob/6355964f89dd1fa6d242ae9b7876baf650cec09e/README.md) +and [contribution rules](https://github.com/Ethical-Tech-CoLab/turnstone/blob/6355964f89dd1fa6d242ae9b7876baf650cec09e/CONTRIBUTING.md). +This is a documentation-level assessment, not a complete metric-engine audit. + +## Recommendation and impact + +Keep the capture pipeline model-free. The project computes pixel-derived +measurements, uses a worst-dimension fitness class, and separates progress +score, missing dimensions and provenance confidence. A language model must +not replace these measurements or live one-instruction guidance. + +An optional **post-capture curator queue** could classify metadata conflicts +or rank already-authorized documentation by relevance to an object. Allow +related-but-not-identical and insufficient evidence. This would affect only +derived review annotations, not the DPA rubric or Plumb geometry. + +Never raise provenance confidence using capture quality. Never use a model to +authorize capture/publication, infer community consent, or change the default +disclosure of sacred/funerary records. Rubric files remain data, never code. + +## Alternatives and test + +Compare a curator checklist and exact metadata rules with a local classifier +first. Jev's [entity-alignment pattern](https://docs.typesafe.ai/cookbooks/entity_alignment.md), +Microsoft-Decision-1 or Sage are possible hosted comparisons only on approved +public records. Their exact hosted checkpoints are not local models. + +Split by object, capture session and source community where appropriate. +Measure false identity links, missing-evidence handling, reviewer time, +calibration and cost per accepted annotation. Include replicas, missing scales, +unmeasured dimensions and source restrictions. Do not use capture labels as +ground truth for authenticity. + +## Gate + +No change to fitness class, score, permissions, reconstruction or live guidance. +Start offline, with reversible annotations and retained source IDs. Stop if +review efficiency is bought by suppressing uncertainty or source-community +control. Any future scoring change requires the repository's conformance tests. +This document proposes a discussion, not a new dependency or runtime feature.